NarraDolma

NarraDolma is a large-scale narrative characterization of Dolma, the 3-trillion-token open pretraining corpus. It contains ~3M passages drawn from ~785K unique documents across all 12 Dolma sub-corpora, each labeled with a fine-grained narrative feature vector produced by NarraBert.

It is the corpus behind the paper Characterizing Narrative Content in Web-scale LLM Pretraining Data and is intended as a resource for studying how narrative qualities are distributed across pretraining data.

• Paper: arXiv:2606.19468
• Collection: Narratives in LLM Pretraining Data
• Source corpus: allenai/dolma

What's in the dataset

Each row is a 3-sentence passage with its Dolma provenance and a 12-feature narrative vector: the 11 NarraBert dimensions plus a derived event-density feature.

Group	Feature	Range	Description
Agency	focalization, emotion, cognition, change_of_state, conflict	1–5	Centrality of each agency feature
Setting	concreteness, temporal_grounding, spatial_grounding, sensory	1–5	How fully the storyworld is realized
Event relations	temporal_sequencing, causal_density	0–1	Fraction of adjacent event pairs that are temporally / causally related
Derived	event_density	≥0	Event triggers per token

Provenance fields: dolma_id (for rehydration), source (Dolma sub-corpus), and topic (WebOrganizer topic, Common Crawl passages only).

How it was built

1. Extraction — ~17M 3-sentence passages sampled from ~5M unique Dolma v1.7 documents, allocated proportionally across sources.
2. Narrative scoring — a DeBERTa binary narrative classifier scores each passage; used to focus sampling on coherent narrative content.
3. Topic classification — Common Crawl passages receive a WebOrganizer topic label (24 topics).
4. Stratified sampling — passages drawn preserving inter-source proportions, with 25% drawn without narrative filtering to keep non-narrative text represented. NarraBert then labels every passage. For analysis, passage vectors can be aggregated to the document level (averaging across a document's passages), yielding one narrative vector per ~785K documents.

Intended use & important caveats

• Relative, not absolute. NarraDolma deliberately over-represents narrative content, so these features describe relative concentration within the sample, not prevalence in raw Dolma. Statements like "X% of source Y is high-interiority" reflect the enriched sample only.
• Event features are noisier. The temporal sequencing and causal density features come from a model that underperforms its LLM teacher on event relations; treat them as higher-noise than the agency and setting features.
• Aggregate to the document level for analysis where possible — passages within a document are not independent.

Rehydration

Each row carries the Dolma unique ID so the original document text can be rejoined from the source corpus.

License & ethical considerations

Released under ODC-By, inherited from Dolma. Dolma is web-scraped and contains toxic, explicit, and personal content. Because NarraDolma pairs passage text with narrative feature labels, it could be misused to retrieve or amplify sensitive personal narratives. It is released for research and auditing; please do not use it to target, extract, or amplify sensitive personal disclosures.

Citation

@misc{johnson2026narrative,
  title         = {Characterizing Narrative Content in Web-scale LLM Pretraining Data},
  author        = {Johnson, Teagan and Ash, Elliott and Piper, Andrew and Antoniak, Maria},
  year          = {2026},
  eprint        = {2606.19468},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2606.19468}
}